GridSearchCV is meant for tuning hyperparameters of a model over some ranges of configurations and parameter values. Like the documentation explains:
https://scikit-learn.org/stable/modules/grid_search.html (and it also has some examples) The (e.g. 10-fold) cross-validation as measure of accuracy (how accurately do different folds attain the value of the statistic) and generalization (that the accuracy remains similar between folds) is at least that what I’m taught at uni. A greater problem is how can one decide, what parameters or e.g. parameter ranges to look for. Since some e.g. float-valued parameters might have some ranges that are “more often used”, while some others that may not work for most of the time. Additionally e.g. the kernels and stuff have some which may have more general robustness, while some others may become computationally very expensive, when combined with some other parameters (such as that in MLPClassifier some activation functions and hidden_layer_sizes may correlate in increased computation cost, while not necessarily increasing accuracy). The best I’ve figured so far is to: Start with few of the most often used / major parameters and try to get them to produce results that are as accurate as possible with still affordable computation time. Only after that consider adding more params. However, I’ve not found much info regarding how the parameters of different methods are ordered in terms of “significance”. One could assume that by the preceding ones are more major than the following ones. However, some of the parameters also clearly “correlate” between each other, so they have cross-effects on accuracy etc. Best is probably just start trying and then perhaps write down, if you notice some general patterns as to what works? There’s also: https://scikit-learn.org/stable/modules/generated/sklearn.pipeline.Pipeline.html for designing “pipelines” or sort of “Design of Experiments” on sklearn algos. Also found this: https://towardsdatascience.com/design-your-engineering-experiment-plan-with-a-simple-python-command-35a6ba52fa35 but have not tried it, nor know if it’s necessary. BR, Matti Lähetetty Windows 10:n Sähköpostista Lähettäjä: lampahome Lähetetty: Thursday, 24 January 2019 11.14 Vastaanottaja: Scikit-learn mailing list Aihe: [scikit-learn] How to determine suitable cluster algo I want to do customized clustering algo for my datasets, that's cuz I don't want to try every algo and its hyperparameters. I though I just define the default range of import hyperparameters ex: number of cluster in K-means. I want to iterate some possible clutering alog like K-means, DBSCAN, AP...etc, and I choose the suitable algo to cluster for me. I'm not sure if that is able to do, but does GridSearchCV work for me? Or any other ways to determine that? thx --- This email has been checked for viruses by Avast antivirus software. https://www.avast.com/antivirus
_______________________________________________ scikit-learn mailing list scikit-learn@python.org https://mail.python.org/mailman/listinfo/scikit-learn
